agentic-server 0.25.0 → 0.25.2

package/esm/router.js

@@ -15,6 +15,37 @@ import { buildProviderHeaders, resolveProvider, resolveUpstreamUrl } from './pro
 import { relayEventStream, withUsageStreamOptions } from './streaming';
 import { transformChatRequest, transformChatResponse, transformEmbedRequest, transformEmbedResponse } from './transforms';
 const log = new Logger('agentic-server');
+/** How long an upstream reason may be before it stops being a message. */
+const UPSTREAM_REASON_LIMIT = 300;
+/**
+ * Name the upstream reason in the message itself.
+ *
+ * Harnesses surface `error.message` and nothing else, so a bare
+ * `LLM provider error: 400` reached a run as an unattributable failure while the
+ * reason — a rejected request shape, a missing key, an unknown model — sat
+ * unread in `error.upstream`.
+ */
+function upstreamErrorMessage(provider, status, body) {
+    const parsed = (() => {
+        try {
+            const error = JSON.parse(body).error;
+            if (typeof error === 'string')
+                return error;
+            const message = error?.message;
+            return typeof message === 'string' ? message : body;
+        }
+        catch {
+            // Not every provider answers an error with JSON; the raw body is the reason.
+            return body;
+        }
+    })().trim().replace(/\s+/g, ' ');
+    const reason = parsed.length > UPSTREAM_REASON_LIMIT
+        ? `${parsed.slice(0, UPSTREAM_REASON_LIMIT)}…`
+        : parsed;
+    return reason
+        ? `${provider.type} provider error ${status}: ${reason}`
+        : `${provider.type} provider error ${status}`;
+}
 export const createRouter = (options) => {
     const router = Router();
     // Metering is backend-agnostic: the caller injects an InferenceSink. When
@@ -90,7 +121,7 @@ export const createRouter = (options) => {
                     });
                 }
                 res.status(upstream.status).json({
-                    error: { message: `LLM provider error: ${upstream.status}`, upstream: text }
+                    error: { message: upstreamErrorMessage(provider, upstream.status, text), upstream: text }
                 });
                 return;
             }
@@ -215,7 +246,7 @@ export const createRouter = (options) => {
                     });
                 }
                 res.status(upstream.status).json({
-                    error: { message: `LLM provider error: ${upstream.status}`, upstream: text }
+                    error: { message: upstreamErrorMessage(provider, upstream.status, text), upstream: text }
                 });
                 return;
             }

package/esm/transforms.js

@@ -1,3 +1,47 @@
+/**
+ * Flatten OpenAI content parts into the single string Ollama's chat api takes.
+ *
+ * Harnesses send `content` as an array of parts — pi always does — and Ollama
+ * answers that with `json: cannot unmarshal array into Go struct field
+ * ChatRequest.messages.content of type string`, which reaches the harness as a
+ * bare `400` and reads as a broken model rather than a dialect mismatch. Image
+ * parts move to Ollama's own `images` field; a part this cannot express fails
+ * loudly rather than being dropped into a prompt the model never sees.
+ */
+function flattenContentParts(messages) {
+    if (!Array.isArray(messages))
+        return messages;
+    return messages.map((message) => {
+        if (!Array.isArray(message.content))
+            return message;
+        const text = [];
+        const images = [];
+        for (const part of message.content) {
+            if (part.type === 'text' || part.type === 'input_text') {
+                text.push(part.text ?? '');
+                continue;
+            }
+            if (part.type === 'image_url') {
+                const url = part.image_url?.url;
+                if (!url)
+                    throw new Error('ollama: an image_url content part carries no url');
+                // Ollama takes base64 bytes, never a fetchable url.
+                const base64 = /^data:[^;]*;base64,(.*)$/.exec(url)?.[1];
+                if (!base64) {
+                    throw new Error('ollama: an image content part must be a base64 data url');
+                }
+                images.push(base64);
+                continue;
+            }
+            throw new Error(`ollama: unsupported content part type '${part.type}'`);
+        }
+        return {
+            ...message,
+            content: text.join('\n'),
+            ...(images.length ? { images: [...(message.images ?? []), ...images] } : {})
+        };
+    });
+}
 export function transformChatRequest(provider, body) {
     let model = body.model || provider.defaultModel;
     if (typeof model === 'string' && model.includes('/')) {
@@ -9,7 +53,7 @@ export function transformChatRequest(provider, body) {
     if (provider.type === 'ollama') {
         return {
             model: model || 'llama3',
-            messages: body.messages,
+            messages: flattenContentParts(body.messages),
             stream: false,
             ...(body.temperature !== undefined && {
                 options: { temperature: body.temperature }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "agentic-server",
-  "version": "0.25.0",
+  "version": "0.25.2",
   "description": "Standalone OpenAI-compatible LLM gateway — multi-provider proxy with pluggable inference metering",
   "main": "index.js",
   "module": "esm/index.js",
@@ -29,13 +29,13 @@
     "test:watch": "jest --watch"
   },
   "dependencies": {
-    "@pgpmjs/logger": "^2.25.0"
+    "@pgpmjs/logger": "^2.25.1"
   },
   "devDependencies": {
     "@types/express": "^5.0.6",
     "@types/node": "^22.19.11",
     "express": "^5.2.1",
-    "makage": "^0.3.0"
+    "makage": "^0.8.0"
   },
   "peerDependencies": {
     "express": "^5.0.0"
@@ -51,5 +51,5 @@
     "metering",
     "constructive"
   ],
-  "gitHead": "e35db12526094b515596ced2dd27a1ec8edd99d5"
+  "gitHead": "5e0bfead2856d1a062b495d343a43e543343f9b8"
 }

package/router.js

@@ -18,6 +18,37 @@ const providers_1 = require("./providers");
 const streaming_1 = require("./streaming");
 const transforms_1 = require("./transforms");
 const log = new logger_1.Logger('agentic-server');
+/** How long an upstream reason may be before it stops being a message. */
+const UPSTREAM_REASON_LIMIT = 300;
+/**
+ * Name the upstream reason in the message itself.
+ *
+ * Harnesses surface `error.message` and nothing else, so a bare
+ * `LLM provider error: 400` reached a run as an unattributable failure while the
+ * reason — a rejected request shape, a missing key, an unknown model — sat
+ * unread in `error.upstream`.
+ */
+function upstreamErrorMessage(provider, status, body) {
+    const parsed = (() => {
+        try {
+            const error = JSON.parse(body).error;
+            if (typeof error === 'string')
+                return error;
+            const message = error?.message;
+            return typeof message === 'string' ? message : body;
+        }
+        catch {
+            // Not every provider answers an error with JSON; the raw body is the reason.
+            return body;
+        }
+    })().trim().replace(/\s+/g, ' ');
+    const reason = parsed.length > UPSTREAM_REASON_LIMIT
+        ? `${parsed.slice(0, UPSTREAM_REASON_LIMIT)}…`
+        : parsed;
+    return reason
+        ? `${provider.type} provider error ${status}: ${reason}`
+        : `${provider.type} provider error ${status}`;
+}
 const createRouter = (options) => {
     const router = (0, express_1.Router)();
     // Metering is backend-agnostic: the caller injects an InferenceSink. When
@@ -93,7 +124,7 @@ const createRouter = (options) => {
                     });
                 }
                 res.status(upstream.status).json({
-                    error: { message: `LLM provider error: ${upstream.status}`, upstream: text }
+                    error: { message: upstreamErrorMessage(provider, upstream.status, text), upstream: text }
                 });
                 return;
             }
@@ -218,7 +249,7 @@ const createRouter = (options) => {
                     });
                 }
                 res.status(upstream.status).json({
-                    error: { message: `LLM provider error: ${upstream.status}`, upstream: text }
+                    error: { message: upstreamErrorMessage(provider, upstream.status, text), upstream: text }
                 });
                 return;
             }

package/transforms.js

@@ -4,6 +4,50 @@ exports.transformChatRequest = transformChatRequest;
 exports.transformEmbedRequest = transformEmbedRequest;
 exports.transformChatResponse = transformChatResponse;
 exports.transformEmbedResponse = transformEmbedResponse;
+/**
+ * Flatten OpenAI content parts into the single string Ollama's chat api takes.
+ *
+ * Harnesses send `content` as an array of parts — pi always does — and Ollama
+ * answers that with `json: cannot unmarshal array into Go struct field
+ * ChatRequest.messages.content of type string`, which reaches the harness as a
+ * bare `400` and reads as a broken model rather than a dialect mismatch. Image
+ * parts move to Ollama's own `images` field; a part this cannot express fails
+ * loudly rather than being dropped into a prompt the model never sees.
+ */
+function flattenContentParts(messages) {
+    if (!Array.isArray(messages))
+        return messages;
+    return messages.map((message) => {
+        if (!Array.isArray(message.content))
+            return message;
+        const text = [];
+        const images = [];
+        for (const part of message.content) {
+            if (part.type === 'text' || part.type === 'input_text') {
+                text.push(part.text ?? '');
+                continue;
+            }
+            if (part.type === 'image_url') {
+                const url = part.image_url?.url;
+                if (!url)
+                    throw new Error('ollama: an image_url content part carries no url');
+                // Ollama takes base64 bytes, never a fetchable url.
+                const base64 = /^data:[^;]*;base64,(.*)$/.exec(url)?.[1];
+                if (!base64) {
+                    throw new Error('ollama: an image content part must be a base64 data url');
+                }
+                images.push(base64);
+                continue;
+            }
+            throw new Error(`ollama: unsupported content part type '${part.type}'`);
+        }
+        return {
+            ...message,
+            content: text.join('\n'),
+            ...(images.length ? { images: [...(message.images ?? []), ...images] } : {})
+        };
+    });
+}
 function transformChatRequest(provider, body) {
     let model = body.model || provider.defaultModel;
     if (typeof model === 'string' && model.includes('/')) {
@@ -15,7 +59,7 @@ function transformChatRequest(provider, body) {
     if (provider.type === 'ollama') {
         return {
             model: model || 'llama3',
-            messages: body.messages,
+            messages: flattenContentParts(body.messages),
             stream: false,
             ...(body.temperature !== undefined && {
                 options: { temperature: body.temperature }